Downscaling for Rainfall Prediction in the Aburrá River Valley and its Supplying Watersheds in Antioquia (Colombia) Using Non-Homogeneous Hidden Markov Models

One of the main issues in projecting climate change scenarios is the coarse scale of global circulation model products.This work employs the statistical downscaling methodology to obtain station-level precipitation series for three climate change scenarios in the region of the Aburrá River and its supplying watersheds, by means of non-homogenous hidden Markov models, using atmospheric variables such as wind, relative humidity, and atmospheric pressure as precipitation predictors. Two approaches are proposed: an annual model and a quarterly one. Each model is trained using precipitation stations to obtain the best number of hidden states that adequately represent the climatology of the area via the BIC criterion. Based on the climatology represented in the states, much better results were obtained with the quarterly model, whose calibrations were used for downscaling the three climate change scenarios. Slight differences were found in the monthly averages, as well as station-scale differences between the probability distributions of daily rainfall for the analyzed scenarios, indicating local alterations in precipitation due to climate change.